| Summary: | 博士 === 大同大學 === 資訊工程學系(所) === 103 === In image retrieval, the well-known SIFT is capable of extracting distinctive features and has been widely used in many fields. However, it is time consuming in matching the features, which slows down the entire process and becomes its major drawback. In the SIFT matching, the Euclidean distance is used as the measurement between two vectors. The calculation of the distance is expensive because it involves the calculation of square of numbers. On the other hand, the scale of the image database usually is too large to adopt linear search for image retrieval. To improve the SIFT matching, this dissertation proposes a fast image retrieval scheme that transforms the SIFT features to binary representation. Accordingly, the complexity of the matching process can be reduced to a much simpler bit-wise operation, which greatly decreases the retrieval time. Furthermore, the proposed scheme utilizes look-up tables (LUT) with four layers of indexes to retrieve similar images. The indexes are derived from the binarized features and can further speed up the retrieval process. Experiments were conducted to examine the usefulness of the binary representation and the LUT, and to demonstrate the effectiveness and efficiency of the proposed scheme. SIFT method and two other methods were also tested for comparison. The experimental results show that the proposed scheme can retrieve images efficiently with comparable accuracy to SIFT and outperforms the other two methods.
|
|---|
